Artificial intelligence-based retinopathy image classification method for diabetic nephropathy patient

Abstract

The invention discloses a diabetic nephropathy patient retinopathy image classification method based on artificial intelligence, and relates to the technical field of medical images, and the method comprises the steps: collecting a retina image, carrying out the denoising and contrast adjustment, and generating a standardized image; carrying out feature constraint by utilizing a generative adversarial network guided by medical priori knowledge in combination with microaneurysm morphology and bleeding area texture features, generating a retina image with enhanced pathological features, and constructing an amplification training data set; training a deep convolutional neural network based on the data set, identifying microaneurysm and bleeding lesion by using a multi-scale feature extraction layer, and outputting a preliminary classification result; and a reinforcement learning environment is constructed in combination with clinician feedback, and a classification decision threshold is dynamically adjusted to optimize lesion grading parameters. According to the method, cross-modal correlation analysis is adopted, retinopathy features and biochemical time sequence data are fused, and the accurate classification capacity of diabetic retinopathy images is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical images, and particularly to an image classification method for diabetic nephropathy patients' retinopathy based on artificial intelligence. Background Art

[0002] In recent years, diabetic retinopathy has become one of the main causes of blindness in adults worldwide. With the development of deep learning technology, medical image analysis methods based on artificial intelligence have made significant progress in the detection and classification of retinopathy. Currently, mainstream AI methods usually rely on large-scale retinopathy image datasets for training to improve the generalization ability of the model. However, due to the limited access to medical data, especially in the group of diabetic nephropathy patients, high-quality and well-annotated datasets of retinopathy are extremely scarce. In addition, the tiny features (microaneurysms, hemorrhages) in retinopathy images are difficult to be accurately identified by traditional computer vision algorithms, which affects the accuracy of disease classification.

[0003] Existing retinopathy image classification methods mainly rely on standard deep learning models, but have the following deficiencies: First, the scarcity of training data leads to insufficient generalization ability of the model when facing different patient groups; Second, existing methods lack a feature constraint mechanism in the identification of microaneurysm and hemorrhage lesions, are easily interfered by noise, and affect the classification accuracy; Third, in the absence of doctor feedback optimization, the classification model may lead to misdiagnosis or missed diagnosis. Therefore, there is an urgent need for an intelligent classification method that can enhance pathological features, optimize classification accuracy, and combine real-time doctor feedback to improve the detection and grading ability of diabetic nephropathy patients' retinopathy. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an image classification method for diabetic nephropathy patients' retinopathy based on artificial intelligence to solve the problems of low classification accuracy of diabetic nephropathy patients' retinopathy images, difficult extraction of lesion area features, and lack of doctor feedback optimization.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides an image classification method for diabetic nephropathy patients' retinopathy based on artificial intelligence, which includes collecting retinopathy images and performing denoising and contrast adjustment to generate standardized images;

[0008] Inputting into a generative adversarial network guided by medical prior knowledge, using a microaneurysm morphological parameter library and a hemorrhage area texture feature map for feature constraint, generating retinopathy images with enhanced pathological features and constructing an augmented training dataset;

[0009] Train a deep convolutional neural network based on an augmented training dataset, use a multi-scale feature extraction layer to identify microaneurysms and bleeding lesions in retinal images, and output an initial diagnosis result including lesion grading;

[0010] Construct a reinforcement learning environment by combining real-time feedback correction data from clinicians, and use a sensitivity-specificity balance factor to dynamically adjust the classification decision threshold to generate optimized grading parameters;

[0011] Load the optimized grading parameters, combine the lesion characteristics of retinal images with the patient's biochemical time-series data, and generate a three-level retinal image classification result for diabetic retinopathy through cross-modal association analysis. As a preferred embodiment of the artificial intelligence-based diabetic nephropathy patient retinal lesion image classification method of the present invention, wherein: the denoising and contrast adjustment refers to,

[0012] Add noise to the retinal image using a Gaussian noise term to generate a noisy image, and label the original retinal image and the corresponding noisy retinal image pair to generate a retinal image dataset;

[0013] Use a U-shaped convolutional neural network for denoising training, and optimize it to minimize the pixel difference between the denoised retinal image and the original retinal image;

[0014] Based on the CLAHE algorithm, divide the denoised retinal image into small blocks, process each small block through histogram equalization to adjust the local contrast, and merge the processing results of the sub-regions to output the retinal image with balanced contrast.

[0015] As a preferred embodiment of the artificial intelligence-based diabetic nephropathy patient retinal lesion image classification method of the present invention, wherein: the generative adversarial network refers to,

[0016] Perform edge detection on the standardized retinal image, extract the edges of microaneurysms by pressing pixel level, extract equidistant sampling points of the microaneurysm lesion contour, retain the low-frequency components to describe the shape features through discrete Fourier transform, and calculate the shape descriptor;

[0017] For the texture of the bleeding area, use a circular neighborhood to calculate the rotation-invariant local binary pattern and capture multi-directional texture features using a differentiable direction encoding layer;

[0018] Based on the morphological features of microaneurysms and the texture features of the bleeding area, the generator adopts a dual-path feature fusion architecture, and the backbone network uses a pre-trained EfficientNet-B4 model to generate a generative adversarial network.

[0019] As a preferred solution of the method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to the present invention, wherein: the multi-scale feature extraction layer includes,

[0020] Three types of convolution kernels arranged in parallel, and dynamically fuse lesion features of different scales through a channel attention mechanism.

[0021] As a preferred solution of the method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to the present invention, wherein: the sensitivity-specificity balance factor includes,

[0022] Dynamically adjust according to the moving window mean rewarded by doctor feedback, and output a threshold adjustment amount through a reinforcement learning Actor-Critic network and evaluate the reinforcement learning state vector to adjust the classification decision threshold.

[0023] As a preferred solution of the method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to the present invention, wherein: the feature splicing and fusion includes,

[0024] Splice the retinal image lesion features and blood glucose time series data in the channel dimension, and use a temporal convolution kernel to extract cross-modal correlation features.

[0025] As a preferred solution of the method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to the present invention, wherein: the construction of the augmented training dataset includes,

[0026] Generate retinal images with enhanced pathological features based on a generative adversarial network guided by medical prior knowledge;

[0027] Perform spatial domain elastic deformation processing on the synthetic images, and the deformation parameters generate random displacement amounts based on a Gaussian distribution to simulate the position and morphological changes of the lesion area.

[0028] As a preferred solution of the method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to the present invention, wherein: generating retinal images with enhanced pathological features includes,

[0029] Perform real-time verification through standard lesion retinal images in the medical knowledge base; when the similarity difference between the microaneurysm morphology or bleeding area texture of the generated image and the standard template is detected, start a dynamic correction mechanism to adjust the dual-path feature fusion weight of the generator to generate enhanced retinal images.

[0030] In a second aspect, the present invention provides a computer device, including a memory and a processor, where the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence as described in the first aspect of the present invention is implemented.

[0031] In a third aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, wherein: when the computer program is executed by a processor, any step of the method for classifying retinal images of diabetic nephropathy patients based on artificial intelligence as described in the first aspect of the present invention is implemented.

[0032] The beneficial effects of the present invention are as follows: A generative adversarial network (GAN) guided by medical prior knowledge is adopted, combined with a microaneurysm morphological parameter library and a bleeding area texture feature map, to generate retinal images with enhanced pathological features, and an augmented training data set is constructed to alleviate the problem of scarce training data. Secondly, a deep convolutional neural network is trained based on the augmented data set, and a multi-scale feature extraction layer is used to improve the recognition ability of micro-lesions, reducing misdiagnosis and missed diagnosis. In addition, a reinforcement learning mechanism is adopted, the classification decision threshold is dynamically adjusted through a sensitivity-specificity balance factor, and the grading parameters are optimized in combination with the real-time feedback of clinicians to enhance the reliability and interpretability of classification. Finally, cross-modal correlation analysis is adopted to fuse retinal lesion features and patients' biochemical time-series data, further improving the accurate classification ability of diabetic retinopathy, and providing a scientific basis for the early diagnosis and personalized treatment of diabetic nephropathy patients. Description of the Drawings

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings based on these drawings without creative efforts.

[0034] Figure 1 It is a flowchart of the method for classifying retinal images of diabetic nephropathy patients based on artificial intelligence in Embodiment 1.

[0035] Figure 2 It is a flowchart of the core process of medical GAN data augmentation in Embodiment 1.

[0036] Figure 3 It is a flowchart of dynamic threshold optimization in Embodiment 1.

[0037] Figure 4 It is a flowchart of the multi-modal fusion process in Embodiment 1. Detailed Embodiments

[0038] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following detailed description of the specific embodiments of the present invention will be made in conjunction with the accompanying drawings of the specification.

[0039] In the following description, many specific details are set forth in order to provide a thorough understanding of the present invention. However, the present invention may be practiced in other ways different from those described herein. Persons skilled in the art may make similar extensions without departing from the spirit of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0040] Secondly, as used herein, an "embodiment" or "embodiments" refers to specific features, structures, or characteristics that may be included in at least one implementation of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or alternative embodiments that exclude each other.

[0041] Embodiment 1, referring to Figures 1 to 4 , this embodiment provides a method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence, including the following steps:

[0042] S1. Collect retinal images and perform denoising and contrast adjustment to generate a standardized image, which includes the following steps:

[0043] S1.1. Use a high-resolution fundus camera to set the parameter resolution to 300 - 600 DPI, select a combination of blue and red light illumination to reduce glare, enhance the visible area of blood vessels and lesions in the image, adjust the exposure time and aperture to ensure uniform image exposure, avoid overexposure or underexposure, and activate the autofocus function to clearly focus on the retinal area and reduce blurring caused by inaccurate focusing.

[0044] Preferably, during the collection process, at least three high-resolution (for example, the image resolution is 20 microns / pixel) retinal images at different angles are taken each time.

[0045] S1.2. Artificially add noise to the retinal image using a Gaussian noise term to generate a noisy image, and label the original retinal image and the corresponding noisy retinal image in pairs to generate a retinal image dataset.

[0046] Among them, the Gaussian noise term is expressed as:

[0047]

[0048] In the formula, I noisy (x, y) is the gray value of the noisy retinal image at the coordinate (x, y), I(x, y) is the gray value of the original retinal image at the coordinate (x, y), (x, y) is the two-dimensional coordinate in the retinal image, N(0, σ 2 ) is the mean of 0 and the variance, ∈ is the noise term and is marked as obeying the normal distribution, is the variance parameter of the noise intensity.

[0049] S1.3. Use a U-shaped convolutional neural network for denoising, and minimize the difference between the noisy retinal image and the true noise-free retinal image by optimizing the L2 loss function.

[0050] Adopt an improved U-Net architecture to denoise the noisy retinal image. During the denoising process, use a four-level encoding-decoding structure. The encoder extracts multi-scale features step by step through convolution and dilated convolution to capture the noise distribution law, and the decoder uses transposed convolution to gradually restore the image details and remove noise interference.

[0051] Among them, the four-level encoding is expressed as:

[0052] DilationRate = [1, 2, 4, 8];

[0053] Among them, the optimized L2 loss function is expressed as:

[0054]

[0055] In the formula, is the original clear retinal image, is the denoised retinal image, N p is the pixel value of the retinal image, and m is the summation index variable.

[0056] S1.4. Based on the CLAHE algorithm (Contrast Limited Adaptive Histogram Equalization algorithm), first divide the input denoised retinal image into small blocks of 8x8 pixels, process each small block through histogram equalization to adjust the local contrast, limit the histogram of each small block after adjustment to avoid over-enhancement, and then merge the processing results of the sub-regions to output the retinal image with equalized contrast.

[0057] Among them, based on the CLAHE algorithm is expressed as:

[0058] I en = CLAHE(I; ClipLimit = 2.0, TileSize = 8×8);

[0059] In the formula, I en is the enhanced retinal image, CLAHE is the Contrast Limited Adaptive Histogram Equalization algorithm, I is the input retinal image, ClipLimit = 2.0 is the histogram clipping value, and TileSize = 8×8 is the local processing block size.

[0060] S1.5. Use RGB color space conversion to convert the retinal image with equalized contrast to the Lab color space, and use the gray world assumption to adjust the color to output the standardized retinal image.

[0061] S2. Input into the generative adversarial network guided by medical prior knowledge, use the microaneurysm morphological parameter library and the hemorrhage area texture feature atlas for feature constraint, generate retinal images with enhanced pathological features, and construct an augmented training dataset, which includes the following steps:

[0062] S2.1. Perform edge detection on the standardized retinal image, extract the edges of microaneurysms at the pixel level, further quantify the shape features of microaneurysm lesions, extract 200 equally spaced sampling points on the microaneurysm lesion contour, expand them into a complex sequence, and calculate the low-frequency component (unit: cycle / mm) that describes the essence of the lesion shape through discrete Fourier transform.

[0063] Among them, extracting 200 equally spaced sampling points on the lesion contour and expanding them into a complex sequence is expressed as:

[0064] z s = u s + jv s ;

[0065] In the formula, z s is the spatial coordinate of the lesion contour, u s is the u coordinate of the s-th sampling point, jv s is the coordinate of jv of the s-th sampling point, and s is the sampling point index.

[0066] Among them, it is calculated through discrete Fourier transform and expressed as:

[0067]

[0068] In the formula, is the index of the frequency component, k = 0, 1, 2,..., 20 is selected in the low-frequency part to quantify the lesion shape features, is the lesion morphological feature of the low-frequency component of the k fieq -th Fourier coefficient, N s is the total number of sampling points, is the complex exponential function, and z n is the complex definition of the contour point coordinates.

[0069] S2.2. Use the rotation-invariant depth LBP operator to model the texture of the hemorrhage area. At the 50μm scale, define a radius annular neighborhood (for example, r = 3 pixels × 20 μm / pixel = 60 μm), and use the differentiable direction encoding layer to capture multi-directional texture features. It is expressed as:

[0070] LBP RI = min{ROR(LBP P , θ)|θ ∈ θ};

[0071] In the formula, LBP RITo extract the texture features of the bleeding area, P is the sampling point of the bleeding area texture, and min is the minimum value of all rotation results {ROR(LBP P , θ)} is the result of LBP P after i rounds of rotation, θ is the set of rotation angles, ROR is the loop operation, and LBP P is the comparison result between the current bleeding area pixel and the neighborhood pixels of the bleeding area.

[0072] S2.3 Based on the morphological features of microaneurysms and the texture feature generator of the bleeding area, a dual-path feature fusion architecture is adopted, and the backbone network extracts multi-scale features through the pre-trained EfficientNet-B4.

[0073] The morphological verification layer is used for connected component analysis to detect the generated features in real time, which is expressed as:

[0074]

[0075] In the formula, ψ circ is the circularity of the lesion area, A is the area of the lesion area, is the logical perimeter of the lesion.

[0076] Among them, feature fusion is expressed as:

[0077] F fusion = α · F low + (1 - α) · F high ;

[0078] In the formula, F low is the edge of the low-level microaneurysm and the texture features of the bleeding area, F high is the high-level lesion feature, F fusion is the fused multi-scale feature vector, and α is the circularity ψ of the lesion area circ dynamically calculated.

[0079] When a pathological morphological structure abnormality is detected, the generator triggers feature re-correction and the adjustment is expressed as:

[0080]

[0081] In the formula, is the shallow edge texture feature of the retinal image, l is the layer index in the generator, is the deep semantic lesion analogy feature of the retinal image, is the Laplacian operator, ⊙ is the element-wise multiplication, is to adjust the weights of the low-level and high-level features.

[0082] Capture the overall image using spectral normalization convolutional layers, and calculate the feature similarity between the generated image and the features in the medical knowledge base within the sliding window, expressed as:

[0083]

[0084] In the formula, is the feature of the retinal image at position (u, v), is the corresponding feature at position (u, v) in the medical knowledge base, (u, v) is the position index, D local (X p ) is the local feature of the retinal generated image, H·W is the height and width of the retinal image.

[0085] S2.4. Construct a composite loss function based on the retinal generated image and the retinal image, introduce the Wasserstein distance to measure the difference in feature distributions, and combine the KL divergence to constrain the texture feature distribution.

[0086] Among them, when constructing the composite loss function, the difference in feature distributions measured by the Wasserstein distance is expressed as:

[0087]

[0088] In the formula, L W is the difference between the retinal image distribution and the retinal generated image distribution, X r is the clinically collected retinal image, P r is the probability distribution of the retinal image, P g is the probability distribution of the retinal generated image, X g is the retinal pathological image synthesized by the adversarial network.

[0089] Among them, the combination of KL divergence to constrain the texture feature distribution is expressed as:

[0090]

[0091] In the formula, L KL is the difference between the texture of the bleeding area and the texture of the generated blood area, p redal (c) is the statistical distribution of the texture of the real bleeding area, p gen (c) is the statistical distribution of the texture of the generated blood area, c is the number of texture categories.

[0092] Use the NSGA-II algorithm (such as the non-dominated sorting genetic algorithm-II) for multi-objective optimization, and dynamically adjust the weights of each objective loss during the training process.

[0093]

[0094] In the formula, w(m) Weight assignment for the m-th target, where M is the total number of targets and m is the static target index R (m) Performance score for the m-th target, where η is the steepness controlling the weight distribution, R (n) Performance score for the current training target, where n is the dynamic target index.

[0095] S2.5. Use a differential equation to simulate the progression of the disease course, incorporate real-time blood glucose data (G(t)) into the generation process, and generate time-series retinal pathological images.

[0096] Among them, the differential equation is expressed as:

[0097]

[0098] In the formula, is the change rate of the lesion diameter, φ(τ) 2 is the simulated time step of the lesion development, the lesion diameter, ∈ is the random growth factor, τ is the time step of the lesion development, 0.025 is the blood glucose influence coefficient, and 0.0018 is the self-limiting coefficient.

[0099] Generate time-series pathological images through numerical integration, output retinal images of the lesion development process, and combine spatial enhancement and time-domain enhancement during data augmentation.

[0100] Among them, spatial domain enhancement is represented by random elastic deformation as:

[0101]

[0102] In the formula, is the spatial deformation of the generated retina, simulating different positions and shapes of the lesion area in the retinal image, σ deform is the displacement standard deviation of the elastic deformation, and Δx and Δy are independent samplings.

[0103] Among them, time-domain enhancement is represented as:

[0104] X interp = αX t +(1 - α)X t+Δt , α ∼ U(0,1);

[0105] In the formula, X t is the state of the retinal lesion at time t, X t+Δt is the state of the retinal lesion at time t + Δt, α ∼ U is the smoothing transition factor for mixing adjacent time steps, and X interp is the intermediate state for simulating the continuous development of the retinal lesion.

[0106] All the lesion images and enhanced images generated by the above method constitute an augmented training dataset.

[0107] S3. Train a deep convolutional neural network by augmenting the training dataset, and use a multi-scale feature extraction layer to identify microaneurysms and bleeding lesions in retinal images, and output an initial diagnosis result including lesion grading.

[0108] S3.1. Extract the enhanced retinal images from the augmented training dataset and divide them into a training set, a validation set, and a test set. Each image is labeled with the locations and lesion types (mild, moderate, severe, etc.) of microaneurysms and bleeding lesions.

[0109] Each input image will be normalized to a fixed size and data normalization will be performed so that the value range of the input data is [0, 1]. In this step, the RGB color space channels of the image will be converted into a format suitable for a multi-layer convolutional neural network.

[0110] S3.2. Based on the deep convolutional neural network, introduce multi-scale feature extraction of retinal image features. During the extraction process, use small-size convolutional kernels (such as 3x3) for feature extraction. In the initial stage of the multi-layer convolutional neural network, extract low-level features (such as edges and textures) of the retinal image through the convolutional layer, and identify microaneurysms and bleeding lesions at different scales.

[0111] To enhance the ability of the trained convolutional neural network to identify lesions of different sizes, use cross-scale convolutional layers (such as 1x1 convolution, 3x3 convolution, and 5x5 convolution) for feature fusion, reduce the size of the feature map through the max pooling layer, retain the feature information of the lesion area, and map the features of the extracted lesion area to the target classification label through multiple fully connected layers.

[0112] S3.3. Use the cross-entropy loss function to optimize the output of the model and use the Adam optimizer to optimize the network. Among them, use the augmented retinal image as the input, perform forward propagation through the deep convolutional neural network to obtain the network output, calculate the loss function, and adjust the network weights.

[0113] S3.4. The trained convolutional neural network outputs the microaneurysm and bleeding areas. At the output layer of the network, use a multi-class softmax function to output the lesion type and its severity level. The output of each lesion area is a three-class, corresponding to mild, moderate, and severe respectively.

[0114] S3.5. Evaluate the trained convolutional neural network by the method of cross-validation, and adjust the structure of the network according to the evaluation results (adjust the learning rate and increase regularization to reduce overfitting).

[0115] After the convolutional neural network is trained, when a new retinal image is input, the model will output the location and type of the lesion area and give the initial diagnosis result of the lesion grade (mild, moderate, severe).

[0116] S4. Combine the real-time feedback of clinicians to correct the data to construct a reinforcement learning environment, and use the sensitivity-specificity balance factor to dynamically adjust the classification decision threshold to generate optimized grading parameters;

[0117] S4.1. Combine with clinicians for real-time feedback, define the doctor's correction action as a discrete reward signal, and input the initial diagnosis result of the lesion grade to construct a time-series clinical feedback data set.

[0118] Among them, defining the doctor's correction action as a discrete reward signal is expressed as:

[0119]

[0120] In the formula, r k is the quantification of the doctor's decision-making behavior.

[0121] Among them, inputting the initial diagnosis result of the lesion grade to construct a time-series clinical feedback data set is expressed as:

[0122]

[0123] In the formula, is the standardized retinal image, is the initial diagnosis result of the lesion grade, and t clin is the timestamp.

[0124] Among them, constructing the reinforcement learning environment is expressed as:

[0125]

[0126] In the formula, S t is the reinforcement learning state vector, P t is the convolutional neural network output probability vector, θ t is the current lesion classification threshold, is the mean reward within the sliding window, is the lesion feature vector of the input image.

[0127] Among them, using the sensitivity-specificity balance factor, define the balance factor as β t ∈[0,1] is expressed as:

[0128]

[0129] In the formula, k = 2 is the adjustment intensity of the sensitivity-specificity balance factor, and μ bmax is the long-term baseline of the sliding window mean.

[0130] When is > μ bmax β(t) → 1, enhancing the sensitivity weight.

[0131] When is < μ bmax β(t) → 0, enhancing the specificity weight.

[0132] Among them, the dynamic adjustment of the classification decision threshold is expressed as:

[0133] R t = β t · Sen t +(1 - β t )· Spe t - λ|θ t - θ t-1 |;

[0134] In the formula, Sen t is the sensitivity weight, Spe t is the specificity weight, and λ = 0.1 is the threshold change penalty coefficient.

[0135] S4.3. Dynamically adjust the classification decision threshold.

[0136] Use the Actor-Critic network to output the threshold adjustment amount and evaluate the reinforcement learning state vector, and update it using policy gradients.

[0137] Among them, the threshold adjustment amount output by the Actor network is expressed as:

[0138] a t = π(S t ; φ) = tanh(W2· ReLU(W1S t + b1)+ b2)

[0139] In the formula, a t is the threshold adjustment action, φ is the Actor network parameter set, W1 is the input layer weight matrix, W2 is the hidden layer weight matrix, ReLU is the hidden layer non-linear transformation, b2 is the hidden layer bias scalar, and tanh compresses the output to [-1, 1].

[0140] Among them, the Critic network evaluates the reinforcement learning state vector and is expressed as:

[0141] V(S t ; ξ) = W c · ReLU(W b S t + b b )+ b c ;

[0142] where \(V(S t )\) is to evaluate the state value of reinforcement learning, \(\xi\) is the parameter set of the Critic network, \(W b \) is the weight matrix of the input layer, \(W c \) is the weight matrix of the hidden layer, \(b b \) is the bias vector of the input layer, \(b c \) is the bias scalar of the hidden layer.

[0143] Among them, and the update using policy gradient is expressed as:

[0144]

[0145] where \) is the classification decision threshold policy gradient, \(J\) is to maximize the long-term reward expectation, \(T\) is the time step window, \) is the reward value calculated by the dynamic balance factor, \) is the variance of the baseline reducing the classification decision threshold policy gradient, and \(N\) is the batch size.

[0146] S4.4. Generate optimized classification parameters.

[0147] Specifically, based on the results of dynamically adjusting the classification decision threshold, use the NSGA-II algorithm to screen the threshold parameters with the optimal classification performance, adopt the Pareto optimal solution generation method, quantify the sensitivity and specificity relationship under different thresholds on the validation set, and construct a trade-off curve.

[0148] Through dynamic-static parameter fusion, integrate the dynamic policy parameters generated by reinforcement learning with the global optimal static threshold to generate deployable optimized classification parameters.

[0149] S5 Load the optimized classification parameters, combine the characteristics of retinal image lesions with the patient's biochemical time series data, and generate a three-level classification result of diabetic retinopathy through cross-modal association analysis.

[0150] S5.1. Based on the generated optimized classification parameters (screened by the NSGA-II algorithm), input the characteristics of retinal image lesions (microaneurysm morphology, hemorrhage texture) and the patient's biochemical time series data (blood glucose concentration, glycated hemoglobin) into cross-modal fusion, extract spatio-temporal association features through a temporal convolutional kernel, and generate retinal lesion fusion features.

[0151] S5.2. For the retinal lesion fusion features, use a deep convolutional neural network (DCNN) to output three classification results (mild, moderate, severe).

[0152] This embodiment also provides a computer device, which is applicable to the situation of the method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence as proposed in the above embodiment.

[0153] This computer device can be a terminal, and this computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of this computer device is used to provide computing and control capabilities. The memory of this computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of this computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a carrier network, NFC (Near Field Communication), or other technologies. The display screen of this computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of this computer device can be a touch layer covered on the display screen, or a button, a trackball, or a touchpad set on the housing of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0154] This embodiment also provides a storage medium, on which a computer program is stored, and when this program is executed by a processor, it implements the method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (abbreviated as SRAM), electrically erasable programmable read-only memory (abbreviated as EEPROM), erasable programmable read-only memory (abbreviated as EPROM), programmable read-only memory (abbreviated as PROM), read-only memory (abbreviated as ROM), magnetic memory, flash memory, a magnetic disk, or an optical disc.

[0155] In summary, the present invention uses a generative adversarial network (GAN) guided by medical prior knowledge, combines the microaneurysm morphological parameter library and the texture feature map of the hemorrhage area, generates retinal images with enhanced pathological features, and constructs an amplified training data set to alleviate the problem of scarce training data. Secondly, a deep convolutional neural network is trained based on the amplified data set, and a multi-scale feature extraction layer is used to improve the recognition ability of tiny lesions and reduce misdiagnosis and missed diagnosis. In addition, a reinforcement learning mechanism is used to dynamically adjust the classification decision threshold through the sensitivity-specificity balance factor, and the classification parameters are optimized in combination with real-time feedback from clinicians to enhance the reliability and interpretability of the classification. Finally, cross-modal association analysis is used to integrate the characteristics of retinal lesions with the patient's biochemical time series data to further improve the accurate classification ability of diabetic retinopathy, providing a scientific basis for the early diagnosis and personalized treatment of patients with diabetic nephropathy.

[0156] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An image classification method for diabetic retinopathy in patients with diabetic kidney disease based on artificial intelligence, characterized in that: including Collecting retinal images, denoising and adjusting the contrast to generate a standardized image; Inputting it into a generative adversarial network guided by medical prior knowledge, using a microaneurysm morphological parameter library and a hemorrhage area texture feature atlas for feature constraint, generating a retinal image with enhanced pathological features and constructing an augmented training dataset; Training a deep convolutional neural network based on the augmented training dataset, using a multi-scale feature extraction layer to identify microaneurysms and hemorrhage lesions in the retinal image, and outputting an initial diagnosis result including lesion grading; Combining the real-time feedback correction data of clinicians to construct a reinforcement learning environment, dynamically adjusting the classification decision threshold using a sensitivity-specificity balance factor, and generating optimized grading parameters; Combining the real-time feedback correction data of clinicians to construct a reinforcement learning environment, dynamically adjusting the classification decision threshold using a sensitivity-specificity balance factor, and generating optimized grading parameters; Loading the optimized grading parameters, combining the lesion characteristics of the retinal image and the patient's biochemical time-series data, and generating a three-level retinal image classification result of diabetic retinopathy through cross-modal association analysis.

2. The method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to claim 1, wherein: The denoising and contrast adjustment mentioned above refers to Adding noise to the retinal image using a Gaussian noise term to generate a noisy image, and annotating the original retinal image and the corresponding noisy retinal image pair to generate a retinal image dataset; Using a U-shaped convolutional neural network for denoising training, and optimizing to minimize the pixel difference between the denoised retinal image and the original retinal image; Based on the CLAHE algorithm, dividing the denoised retinal image into small blocks, processing each small block through histogram equalization to adjust the local contrast, and merging the processing results of the sub-regions to output a retinal image with balanced contrast.

3. The method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to claim 1, characterized in that: The generative adversarial network mentioned above refers to Performing edge detection on the standardized retinal image, extracting the edges of microaneurysms by pressing pixel levels, extracting equidistant sampling points of the microaneurysm lesion contour, retaining the low-frequency components to describe the shape features through discrete Fourier transform, and calculating the shape descriptor; For the hemorrhage area texture, using a circular neighborhood to calculate the rotation-invariant local binary pattern and using a differentiable direction encoding layer to capture multi-directional texture features. Based on the morphological features of microaneurysms and the texture features of the hemorrhage area, the generator adopts a dual-path feature fusion architecture, and the backbone network uses a pre-trained EfficientNet-B4 model, which is a generative adversarial network.

4. The method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to claim 1, characterized in that: The multi-scale feature extraction layer includes Three types of convolutional kernels set in parallel, dynamically fusing lesion features of different scales through a channel attention mechanism.

5. The method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to claim 1, characterized in that: The sensitivity-specificity balance factor includes Dynamically adjusted according to the sliding window mean of the doctor feedback reward, and adjusting the classification decision threshold by outputting the threshold adjustment amount and evaluating the reinforcement learning state vector through a reinforcement learning Actor-Critic network.

6. The method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to claim 1, wherein: The feature stitching and fusion includes Stitching the lesion features of the retinal image and the blood glucose time-series data in the channel dimension, and using a temporal convolutional kernel to extract cross-modal association features.

7. The method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to claim 1, wherein: The construction of the augmented training dataset includes Synthesizing a retinal image with enhanced pathological features based on a generative adversarial network guided by medical prior knowledge; Perform spatial domain elastic deformation processing on the synthetic image, where the deformation parameters generate random displacement amounts based on a Gaussian distribution to simulate the position and morphological changes of the lesion area; Perform time domain interpolation enhancement on the time series pathological images, generate intermediate states by randomly linearly mixing the images of adjacent time steps, and construct an augmented training dataset.

8. The method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to claim 1, characterized in that: Generating a pathologically enhanced retinal image includes Performing real-time verification through standard lesion retinal images in the medical knowledge base; when the similarity difference between the microaneurysm morphology or the texture of the bleeding area of the generated image and the standard template is detected, start the dynamic correction mechanism to adjust the dual-path feature fusion weights of the generator and generate an enhanced retinal image.

9. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the method for classifying retinal lesion images of diabetic nephropathy patients based on artificial intelligence according to any one of claims 1 to 7.

Images